From a legacy-code necromancer to a test fraud auditor. Questionable personalities. Very specific engineering jobs.
Meet the team
Call one when you need them. Each has a specific job and a stopping point.
“Their reasons didn’t leave with them.”
Trace the purpose of suspicious legacy code through current callers, tests and Git history.
$necromancer
Can we remove this compatibility branch?Even “it’s the cache” needs evidence. Use experiments to distinguish causes.
Fail before. Pass after. Keep evidence from actual execution.
One button doesn’t need a new architecture. Finish within the requested scope.
Good results and adverse results are both public. General efficiency gains for the whole team remain unproven.
Five new tasks, two arms, one fresh session per cell. See the gains and the increased costs together.
mother-in-law-confirmation-2026-09-122026-09-12 / e76efcf
With skill 93.2% / Without skill 100%
With skill 71.3% / Without skill 100%
Without skill → With skill · False positives 0 → 0
Equal-weight mean of per-task skill/no-skill ratios. This is neither a reduction in summed tokens nor a billing saving.
Sum of input + output across every model response. Example: 10,000 tokens on the first response + 12,000 on the next = 22,000 for the task. Conversation and tool results sent again count in the next input. Cached input is already included in each response’s input; this is neither billing nor final-answer length.
Tokens decreased on 3, increased on 2 / 5 tasks
100% is the no-skill usage for that task. A longer bar means more resource use. Read the actual usage and increase or decrease alongside each task.
GPT-6 Astra · medium · 5 new tasks × 2 arms × 1 fresh session
| Task | Arm | Cumulative tokens | Time (s) | Target |
|---|---|---|---|---|
Clear addressaddress-clear-boundary | Without skill | 64,055 | 56.717 | Met |
Clear addressaddress-clear-boundary | With skill | 83,165 | 32.789 | Met |
Guarded searchguarded-profile-search | Without skill | 80,127 | 52.006 | Met |
Guarded searchguarded-profile-search | With skill | 66,061 | 25.534 | Met |
Double submitinvoice-double-submit | Without skill | 99,730 | 71.950 | Met |
Double submitinvoice-double-submit | With skill | 68,259 | 59.735 | Met |
Stale errorlookup-stale-error | Without skill | 83,396 | 91.347 | Met |
Stale errorlookup-stale-error | With skill | 66,013 | 36.634 | Met |
Nested city searchnested-city-runner | Without skill | 82,009 | 45.383 | Met |
Nested city searchnested-city-runner | With skill | 87,217 | 57.357 | Met |
| Raw sums | Without skill | 409,317 | 317.403 | 5/5 |
| Raw sums | With skill | 370,715 | 212.049 | 5/5 |
Total tokens include cached input + output. Ratios of raw sums differ from the equal-weight task means plotted above.
Mixed whole-team results and adverse outcomes remain available in the decision index below.
Evidence published on GitHubApplies to the versions named there · may differ from this page’s download Featured comparisonThe experiment and limitations linked by featured.jsonSkills guide the workflow. Results depend on the model, tool access, project and user instructions.
Have Node.js/npm and Git? You’re ready. Choose your skills and supported agents.
npx skills add SoonGwan/questionable-hiresNo background processes · No telemetry
Runs the independent skills CLI. Review the CLI and selected skills before execution.
Eight skills, including Receipt v3. This snapshot differs from the GitHub release installed by the command above.